Before the Prompt:
The Upstream Craft

If episode one was the diagnosis, this is where we get to work. The single biggest lever for getting useful output from an AI tool is not the prompt. It is the thinking that happens before the prompt — the upstream craft of framing a problem, decomposing it cleanly, surfacing the assumptions inside it, and stress-testing the whole structure before a single character is typed into the model. This is unglamorous work. It is also the work that almost nobody is doing well right now, and the work that quietly determines whether everything downstream is useful or hollow.

Let me show you why it matters. Then let me show you how to do it.

The cost of skipping the upstream work

Imagine two analysts, both asked the same Monday-morning question by their CEO. The question is one I have actually heard a CEO ask, almost verbatim — "are we losing pricing power in our commercial book?"

Analyst A opens the AI tool immediately. They paste the question, ask for a structured response, get a tidy three-page answer with bullet points on competitor pricing, market share trends, channel feedback and reinsurer signals. They polish the output. They send it up. They feel productive. The whole thing took ninety minutes.

Analyst B does not open the AI tool yet. They spend forty-five minutes with a notebook. They ask: what does "losing pricing power" actually mean here? Is it competitor behaviour, our own underwriting discipline, channel pressure, reinsurance terms — or all four? Over what time horizon — last quarter, last year, the cycle? In which segments — fire, engineering, marine, liability? Compared to what — our own history, our peers, the tariff regime? They write down twelve such questions. Then they pick the four that will actually move the CEO's thinking — that the CEO does not already know the answer to, and that, when answered, will change a decision he is currently weighing. Only then do they open the AI tool — and they ask four sharp, bounded questions instead of one vague one. Total time, including AI work — about three hours.

Both analysts will produce a memo. Analyst A's will sound complete. Analyst B's will be complete. The difference will be visible in the very first board question. Analyst A will produce an answer the CEO already had. Analyst B will produce one he did not.

The prompt is the visible part of the work. The framing is the part that decides whether the work is any good. Almost everyone is optimising the visible part.

The four moves that follow — frame, decompose, assume, pre-mortem — are how you become Analyst B. They are not a framework I invented. They are the working discipline of every senior strategist I have ever respected, drawn from McKinsey's old issue-tree tradition, from Barbara Minto's pyramid principle, from Gary Klein's pre-mortem research, and from the underwriter's habit of always asking what the policy actually covers before quoting on it. None of this is new. What is new is that the AI era now punishes people who skip these moves, in a way it did not before.

Move one — Frame the actual question

Most business questions arrive in vague form. Should we enter this market? Are we competitive? What is our digital strategy? Are we losing pricing power? These are not questions an AI can answer well — because they are not questions any human could answer well either. They are starting points. Invitations to do the framing work. The mistake is treating them as the question itself.

The framing move has three components, and you have to do all three.

First, find the decision. Behind every vague question is a specific decision someone is trying to make. "Should we enter this market" might really be "should we commit Rs 50 crore of distribution capex to a Tier-2 expansion in FY27." That is a question you can actually answer. "Are we losing pricing power" might really be "should we change our minimum-rate discipline at the underwriting floor for fire risks above Rs 100 crore sum insured." Behind every fuzzy question, find the sharp decision. If you cannot find it, go ask the person who asked the question what decision they are weighing. Most CEOs will be relieved you asked.

Second, find the audience. The same question answered for a board, an investment committee, a regulator, and an internal exco looks completely different. The board wants strategic implications and risk. The investment committee wants capital deployment logic. The regulator wants prudential evidence. The exco wants the operational plan. One AI prompt cannot serve all four audiences at once, and yet I see analysts ask AI for "an analysis" without ever defining who will read it. The result is generic, because the audience was generic.

Third, find the time horizon. A pricing question over the next quarter is a tactical analysis. The same question over a five-year cycle is a strategic one. They use different evidence, different logic, different conclusions. Most AI prompts fail to specify the horizon, and the model defaults to whatever is statistically most common in its training data — which, for business questions, tends to land in a vague middle distance that suits no real decision.

Until you have answered those three — decision, audience, horizon — you should not be talking to the AI. You should be talking to yourself, or to the person who asked the original question. This is the cheapest, fastest, most leveraged move in the entire upstream stack, and almost nobody does it deliberately.

Move two — Decompose ruthlessly

Once the question is framed, the next move is decomposition. This is where the consulting world's old MECE discipline still earns its keep — break the question into parts that are mutually exclusive (no overlap) and collectively exhaustive (nothing missing). Done well, decomposition turns one impossible question into a set of small, tractable ones, each of which the AI can actually help with.

Take "are we losing pricing power in our commercial book." A clean decomposition might split it into four branches.

Market dynamics — what are competitors and reinsurers actually doing on rate, terms and capacity? Portfolio mix — are we writing the same business at lower rates, or different (riskier) business at the same rates? Channel behaviour — are intermediaries forcing concessions, and are they concentrated in particular segments? Internal discipline — have our own underwriting standards drifted, and if so, where and how much? Each of those four splits further. Channel behaviour might split into broker behaviour, agent behaviour, direct/digital behaviour. Internal discipline might split by segment, by underwriter, by minimum-rate compliance. By the time you are done, you have a logic tree with maybe twenty sub-questions — each of which is small enough that an AI prompt can address it usefully.

This is where AI becomes genuinely useful. The model is excellent at answering bounded sub-questions. It is bad at answering meta-questions. Decomposition is what turns the second into the first. A logic tree is, in effect, the rough outline of every prompt you will run for the next two days.

One more point on decomposition. The hardest part is making your decomposition genuinely mutually exclusive. Most first attempts have overlap. "Market dynamics" and "channel behaviour" can both contain reinsurance pressure if you are not careful. Spending an extra fifteen minutes cleaning up the overlap before you start prompting is worth more than the next four hours of analysis. Overlap in the tree means the same evidence will get used twice in the final memo, and a sharp reader will catch it.

Move three — Surface the assumptions

Every analysis carries assumptions. The dangerous ones are the invisible ones — the ones the analyst never wrote down, possibly never even noticed they were making. Before you reach for the model, write down what you are taking for granted.

For the pricing-power example, the assumptions might run like this. I am assuming the regulatory regime stays stable. I am assuming the soft cycle continues for at least 18 months. I am assuming our reinsurance terms hold at next renewal. I am assuming the competitor behaviour we are seeing is rational rather than driven by short-term capital pressure. I am assuming our own underwriting discipline is being measured correctly by our internal MIS. Each of those is an assumption. Each could be wrong. If any of them is wrong, the conclusion of the analysis changes.

Now do the harder thing. Rank those assumptions by two dimensions — how confident are you, and how much does the answer change if you are wrong. The assumptions in the high-impact-low-confidence quadrant are the ones the entire piece of analysis hinges on. They deserve their own analysis. The model will not flag them for you. You have to flag them yourself, and then ask the model to test them.

This is where most AI-driven work falls apart on contact with a sharp board. The presenter has not separated the conclusions from the assumptions they rest on. When a director pulls on a single assumption, the whole stack unravels — because the writer did not know which assumption was load-bearing in the first place. The five minutes spent ranking your assumptions before you prompt is the five minutes that saves you from collapse during Q&A.

Move four — Pressure-test before you prompt

The last upstream move is the pre-mortem. The technique is from Gary Klein at Klein Associates, originally developed for military and emergency-services planning. It is the simplest cognitive trick you will ever use. Imagine the board has just heard your final memo. Imagine they hated it. Why?

Write down the three sharpest objections a thoughtful sceptic would raise. Maybe the data is too thin. Maybe the comparison set is wrong. Maybe a major risk has been ignored. Maybe the conclusion is right but for reasons you have not articulated. Maybe the recommendation works on paper but fails on contact with the actual organisation — which lacks the talent, the system, or the political space to execute it. Write down each objection in the sceptic's voice. Make it sting.

Each of those objections then becomes a directed prompt to the AI. "What is the strongest counter-argument to the claim that we are losing pricing power because of channel pressure rather than internal discipline drift?" is a far more useful prompt than "give me a competitive analysis." The model is excellent at adversarial framing if you tell it to be adversarial. It will not volunteer that posture — its default is to be helpful, agreeable and on-side. The pre-mortem flips that default. It forces the model to argue against you, which is exactly what you need to find the weak parts of your own structure.

The pre-mortem is also where humility lives. If you genuinely cannot think of three sharp objections to your own emerging conclusion, you do not yet understand the problem well enough. Either go back to decomposition and find branches you have not yet explored, or go talk to someone who disagrees with you. The absence of a counter-argument is not strength. It is blindness.

Now — and only now — do you prompt

By the time you actually open the AI tool, the work has already paid for itself. You know the decision, the audience, the time horizon. You have a logic tree with maybe twenty sub-questions. You have a list of assumptions ranked by impact and confidence. You have three pre-mortem objections. Your prompts can now be sharp, bounded, and aimed at specific gaps in your thinking — not vague gestures at a topic.

A prompt at this stage might read like this. "For an Indian general insurance commercial book, find me three pieces of public evidence on how channel commission economics in fire and engineering have moved since the de-tariffication of April 2024. I am particularly interested in evidence that distinguishes broker behaviour from agent behaviour. Be specific about sources and time periods. If the evidence is thin, say so." That is a prompt that will get you something useful. Compare it to "give me a view on commercial pricing in Indian general insurance." Same topic. Wildly different output.

This is the moment AI starts behaving like the genuine accelerant it can be. You are no longer asking it to think for you. You are asking it to extend your thinking — find data you do not have, run a counter-argument you have not run, structure a sub-analysis you have already framed. The output will still need scrutiny — that is the subject of the next essay — but it will be ten times more useful than the output of an unstructured prompt, because the input was structured.

The discipline is the differentiator

Here is what I have noticed about the people who do this well. They are not necessarily the smartest in the room. They are the ones who have built the habit of doing the upstream work — every time, even when the deadline is tight, even when the question seems simple, even when the AI tool is sitting open and inviting them to skip ahead. The discipline is the differentiator. Not the talent.

And it compounds, in exactly the sense that gives the first MacSays series its name. Every time you do the upstream work, you build the muscle. You also build a personal library of well-framed problems, well-decomposed trees, and well-tested assumptions that you can draw on next time. Every time you skip it, you let the muscle atrophy and you accept worse output. After a year, the gap between the two professionals is enormous — and they may have started in roughly the same place.

One last thing. The upstream work cannot be outsourced to the AI itself. I know — that is the obvious next thought. "Can I just ask the AI to frame the problem for me?" You can, and the output will look reasonable, and it will be subtly wrong in ways you will not notice until the board calls them out. The AI does not know your firm, your team, your political constraints, your customer history, your unspoken realities. The framing has to come from you, because it has to be grounded in context the model does not have. The upstream work is the irreducibly human part of the entire stack. Which is exactly why it is the part that matters most.

In the final essay, I will turn to the downstream craft — what to do after the model gives you an answer. How to interrogate it, integrate it, and make the resulting work genuinely yours. That is where the second half of the value lives, and it is where a different set of disciplines apply.